FFAM: Feature Factorization Activation Map for Explanation of 3D Detectors

Summary

This paper proposes a feature factorization activation map to explain 3D detectors. It uses non-maximum matrix factorization (NMF) to obtain a global concept activation map and then refine it with feature gradients of an object-specific loss. A voxel upsampling strategy is further proposed to upsample sparse voxels to align the granularity between the activation map and input point cloud. Both quantitative and qualitative results validate the efficacy of the proposed method.

Strengths

- The basic idea is easy to follow, and the comparison with previous works (both in 2D and 3D) is clear. - The key motivation and contributions are clearly presented. - The content of methodology and implementation are well organized. - The visualization results are impressive. Both the quantitative and qualitative results convincingly support the efficacy of the proposed explanation method.

Weaknesses

The only concern is about the application, or the value in applications, of this proposed method. I understand this is a good adaptation and attempt to apply such explanation methods on LiDAR-based 3D detectors, but it is still unclear how it can benefit the procedure of improving 3D detectors or downstream applications. I would be more interested in how it produces a stronger 3D detector or a 3D detector with better controllability and safety, especially for safety-critical scenarios like autonomous driving. In addition, there are also related areas or applications, such as 3D detection in indoor scenes (ScanNet, SUN RGB-D), etc, and camera-based/multi-modality 3D detection. I am also curious about the performance of such explanation methods when applied to those problems. How about the performance, and is there any new challenge?

Questions

None.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

None.

Summary

The paper proposes a method called Feature Factorization Activation Map (FFAM) to provide visual explanations for 3D object detectors based on LiDAR data. This method addresses the interpretability issue in 3D detectors by using non-negative matrix factorization to generate concept activation maps and refining these maps using object-specific gradients. The approach is designed to handle the unique challenges of 3D point cloud data, such as sparsity and the need for object-specific saliency maps. The paper evaluates FFAM against existing methods and demonstrates its effectiveness through qualitative and quantitative experiments.

Strengths

1. Using non-negative matrix factorization to generate concept activation maps is novel and well-justified for the application. 2. The method is evaluated on multiple datasets and compared against state-of-the-art methods, demonstrating its superiority in producing high-quality visual explanations. 3. The paper provides a clear and detailed description of the methodology, including the feature factorization, gradient weighting, and voxel upsampling processes. 4. The proposed method has practical implications for improving the interpretability of 3D object detectors, which is crucial for applications in autonomous driving and robotics.

Weaknesses

1. The method involves several computationally intensive steps, such as non-negative matrix factorization and voxel upsampling, which may limit its applicability in real-time systems. 2. The paper focuses primarily on LiDAR-based 3D detectors. Discussing how the method could be adapted or extended to other types of 3D data or detection systems would be beneficial. 3. While the evaluation is comprehensive, it primarily focuses on two datasets (KITTI and Waymo Open). Additional datasets and scenarios could further validate the robustness and generalizability of the method.

Questions

1. Real-Time Applicability: How does the computational overhead of FFAM compare to existing methods in real-time applications, especially in autonomous driving scenarios? 2. How sensitive is FFAM to the choice of detector? We noticed that two kinds of detectors were used in this paper. More comparison and analysis are necessary. 3. Can FFAM be extended to other 3D detectors or modalities, such as RGB-D sensors or radar data? If so, what modifications would be necessary?

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

1. FFAM requires access to the feature maps within 3D detectors, which may not always be possible, especially for proprietary or closed-source systems. 2. The method's scalability to large-scale or real-time applications is not fully addressed. The computational requirements may be prohibitive for some practical applications. 3. The method is tailored to LiDAR data and may not directly translate to other types of 3D data without significant modifications.

Summary

This paper addresses the challenge of explanation and interpretability in 3D detection methods. It introduces a Feature Factorization Activation Map (FFAM), which utilizes non-negative matrix factorization (NMF) and object-specific gradient weighting to generate global and object-specific activation maps at the voxel level. An up-sampling method is subsequently employed to produce per-point activation maps. Extensive quantitative and qualitative experiments demonstrate the effectiveness of the proposed FFAM in generating saliency maps for various points over the previous method.

Strengths

- The topic is interesting and warrants investigation. As discussed in Section 4.3, the research can significantly enhance 3D object detectors, particularly in identifying false positive modes. - The quantitative results are convincing and relevant to the topic. - Overall, the paper is well-written and easy to comprehend. - The code is made available, promoting transparency and reproducibility. - The paper discusses the limitation of the proposed method, specifically the necessity of accessing the feature map.

Weaknesses

- The rationale for using NMF is unclear. It appears to be a learning-based, PCA-like method to extract saliency from voxel features. How does it compare to the proposed method on the global activation map? - There is a lack of ablation studies, such as those examining the parameter 𝛾 - The gradient weighting method is intriguing and seems sensible. Is this method original to your work? - It would be beneficial to use 𝑊 as a weight in Equation 2. The current version is a little confusing. - In Section 4.1, it is claimed that the saliency map generated by FFAM is superior to occAM because it is more focused on the object. Please provide further validation on why a clearer activation map is considered better.

Questions

Please check weaknesses.

Rating

5

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

It would strengthen the paper if an example could be provided demonstrating how FFAM can help identify some non-trivial error modes and offer insights for improving the detector.

Summary

The paper proposes a “FFAM” for visual visualization of 3D Detectors. It introduces a non-negative matrix factorization (NMF) to decomposing 3D features into the product of two non-negative matrices. Besides, an object-specific loss is utilized to generate the object-specific saliency maps. Finally, the voxel upsampling is used to recover the resolution of the activation maps.

Strengths

1. This work introduces NMF in explaining point cloud detectors and utilizes feature gradients of an object-specific loss to generate object-specific saliency maps. 2. A voxel upsampling strategy is proposed to upsample sparse voxels.

Weaknesses

1. The description of Non-negative Matrix Factorization (NMF) is unclear. When the point cloud is large in scale, NMF may become unstable and exhibit a long convergence time. 2. The core innovation of this paper lies in feature factorization. The authors directly chose NMF but did not provide a reason for this choice. For instance, Principal Component Analysis (PCA) can also achieve feature factorization. 3. The object-specific gradient is element-wise multiplied with the global concept activation map to obtain the specific activation map of the object. In other words, the object-specific gradient inhibits the activation that do not belong to the current object. Therefore, it seems reasonable to directly use the weighting of the object-specific gradient as the activation maps. 4. Since the gradient map G can already represent the current object, why not perform NMF on the gradient map and then use a method similar to obtaining a global activation map to derive the object activation map? 5. Object-Specific Gradient Weighting is a general module. Can it be applied to other methods of generating activation maps, such as OccAM? 6. Regarding voxel upsampling, the authors provide a limited description. I am curious about why they chose the Gaussian kernel and how it compares to other upsampling methods such as trilinear interpolation and transpose convolution.

Questions

Please the weakness.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors mentioned the limitations and societal impact in the checklist.

Final Decision

Thanks for the author's efforts in addressing my questions. Given the response, I can better understand the application and adaptation for other scenarios, but I think there are still limited new insights regarding how the proposed method can help improve current algorithms. The connection between the analysis and the proposed method is a little weak. Hence, I cannot raise my rating and would keep the original "borderline acceptance" recommendation.